FDRN: A fast deformable registration network for medical images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deformable image registration is crucial in medical imaging but computationally intensive.
• Conventional methods often compromise accuracy for speed.
• There is a need for efficient and accurate deformable registration techniques.

Purpose of the Study:

• To propose a fast unsupervised convolutional neural network (FDRN) for deformable image registration.
• To improve both the accuracy and runtime of volumetric image registration.
• To address the limitations of traditional iterative registration methods.

Main Methods:

• FDRN utilizes a compact encoder-decoder architecture taking fixed and moving images as input.
• It outputs a 3D displacement vector field (DVF) for voxel mapping.
• Key innovations include additive forwarding, deep supervision with a coarse-to-fine strategy, and an optional multi-label segmentation loss for enhanced accuracy.

Main Results:

• FDRN demonstrated superior performance compared to state-of-the-art methods like VoxelMorph.
• Achieved a 1.46% higher Dice score on the LPBA40 dataset.
• Outperformed other methods in Dice score and NCC on multiple unseen MRI datasets (CUMC12, MGH10, ABIDE, ADNI).

Conclusions:

• FDRN offers enhanced performance in brain MRI registration due to its efficient autoencoder structure and learning strategy.
• The proposed method generalizes well across different medical image types and anatomies.
• FDRN represents a significant advancement in fast and accurate deformable image registration.